Clinical Evidence

Peer-reviewed results and before-and-after comparisons for SwiftMR® and SwiftSight-Brain.

Denoising

Improve SNR with well-preserved tissue contrast

Super-resolution

Sharper lesion and structural conspicuity

Acceleration

Shorter scan time, less motion artifact

Artifact reduction

Cleaner truncation and susceptibility artifacts

MS lesion tracking

New, enlarging, and stable lesions across exams

Diagnostic Performance

Peer-reviewed studies that measured diagnostic performance in the same patients with and without SwiftMR.

Cervical spine imaging

47% faster scans, with higher sensitivity for foraminal stenosis

Long scans are hard on patients with spinal disease and invite motion artifacts. Each patient had a conventional scan and one 47 to 48% shorter with SwiftMR on the same day, read by four radiologists.

Sensitivity for neural foraminal stenosis: Conventional, 65%. With SwiftMR, 73%. Difference 7 points (95% CI 2–13), p = .005.

• 50 patients
• Prospective
• 4 readers

Koo et al. AJNR, 2025

Brain tumor imaging

More accurate glioma grading from DCE perfusion maps

DCE-MRI trades signal-to-noise for temporal resolution, which adds variability to perfusion measurements. With SwiftMR, mean extracellular volume (Ve) separated low- from high-grade diffuse glioma more accurately.

AUROC for WHO grade, mean extracellular volume: Standard DCE, 0.83 (95% CI 0.76–0.90). With SwiftMR, 0.88 (95% CI 0.83–0.94). p = .02.

• 306 patients
• Retrospective
• Histopathology reference

Lee et al. Scientific Reports, 2024

Temporal lobe epilepsy

More accurate TLE diagnosis on thin‑slice MRI

Thin slices show the hippocampus in more detail but add noise. With SwiftMR, two neuroradiologists diagnosed TLE more accurately from the same 1.5 mm coronal T2-weighted images.

Pooled diagnostic accuracy, 1.5 mm images: Without SwiftMR, 73.1% (95% CI 65.5–79.5). With SwiftMR, 80.8% (95% CI 74.0–86.1). p = .010, pooled across both readers.

• 117 patients
• 2 blinded readers

Suh et al. Korean Journal of Radiology, 2024

Intracranial vessel wall imaging

Three times the sensitivity for culprit plaques

Limited signal-to-noise makes plaque features hard to see, especially in small distal arteries. Using intraplaque hemorrhage, sensitivity for culprit plaques rose from 23.5% to 70.6% with SwiftMR, and specificity held at 94.3%.

AUROC for culprit plaque, intraplaque hemorrhage: Conventional, 0.59 (95% CI 0.48–0.70). With SwiftMR, 0.82 (95% CI 0.71–0.94). p < .001, DeLong test.

• 70 plaques

Seo et al. Scientific Reports, 2024

Values as published, with 95% confidence intervals (whiskers) where the study reported them. AUROC bars start at 0.5, the value for a test no better than chance. Each study compared images of the same patients with and without SwiftMR.

Improved SNR with well-preserved tissue contrast, revealing structures that noise obscures, without changing your scan protocol.

Vessel wall noise reduction

5:22 — Without SwiftMR

5:22 — With SwiftMR

C-Spine T2 STIR AX

1:30 — Without SwiftMR

1:24 — With SwiftMR

Deep learning improves quality of intracranial vessel wall MRI for better characterization of potentially culprit plaques

SwiftMR provided significant improvement in SNR and spatial resolution in vessel wall imaging, leading to detection of vessel wall lesions and intraplaque hemorrhage compared to conventional image.

Seo et al.. Scientific Reports. Aug 2024.

Improved lesion and structural conspicuity, enabling thin-slice imaging that would otherwise be limited by SNR constraints.

1:19 — Without SwiftMR

1:19 — With SwiftMR

C-Spine T2* MEDIC AX

4:21 — Without SwiftMR

2:45 — With SwiftMR·37% faster

Improving diagnostic performance of MRI for temporal lobe epilepsy with deep learning-based image reconstruction in patients with suspected focal epilepsy

SwiftMR resulted in significantly improved resolution and structural conspicuity in hippocampal imaging compared to routine imaging. Hippocampal striation blurring was suspected on routine 3mm MRI, but became more obvious on 1.5mm MRI + SwiftMR.

Suh et al.. Korean Journal of Radiology. 2024;25(4):374-383.

Shorter scan time means less patient movement, minimizing chances of motion artifacts, directly improving image quality as a secondary benefit.

L-Spine T2 TSE SAG

2:12 — Without SwiftMR

1:22 — With SwiftMR·38% faster

Knee PDWI FS

2:42 — Without SwiftMR

1:36 — With SwiftMR·41% faster

Deep learning-based reconstruction for acceleration of lumbar spine MRI: a prospective comparison with standard MRI

Prospective comparison showing 32.3% average acquisition time saving versus standard protocol, with no degradation in image quality or diagnostic performance.

Yoo et al.. Eur Radiol. 2023;33(12):8656-8668.

Highly accelerated knee magnetic resonance imaging using deep neural network (DNN)-based reconstruction: prospective, multi-reader, multi-vendor study

Prospective, multi-reader, multi-vendor study demonstrating average scan time reduction of 41% using SwiftMR, without degrading image quality or diagnostic performance across readers or scanner platforms.

Lee et al.. Sci Rep. 2023;13:17264. PMID: 37828048.

Artifact Reduction

Improve truncation and susceptibility artifacts, producing cleaner images from the same acquisition without reacquisition or protocol changes.

Susceptibility artifacts reduction

2:09 — Without SwiftMR

2:09 — With SwiftMR

Geometric distortion improvement

Brain DWI b1000 · 2D EPI Diffusion · Same acquisition 3.0T Siemens MAGNETOM Skyra

1:30 — Without SwiftMR · 1.5×1.9×3.0 mm

1:30 — With SwiftMR · 1.5×1.9×3.0 mm

Truncation artifact improvement

Brain CE T1WI

MS Lesion Tracking

Longitudinal MS tracking compares exams taken months or years apart, often on different scanners. A change in vendor, field strength, or protocol can alter noise and resolution, so automated measurements can shift even when the patient has not changed.

SwiftSight-Brain minimizes cross-scanner variation by normalizing and improving image quality before it compares exams, then classifies each lesion as new, enlarging, or stable.

Below are three examples of real-world variability between exams: a different scanner model, a different field strength, and a different vendor. In each, SwiftSight-Brain tracks the lesions consistently despite the change.

Cross-scanner variation in brain volumetry

Largest difference in hippocampal volume for one subject scanned on three scanners.

Across scanner models

Retrospective case · Female patient, age 31–39 Siemens MAGNETOM Skyra 3T and Vida 3T

• New
• Enlarging
• Stable

This patient's baseline at age 31 was acquired on a Siemens Skyra 3T, and the follow-up eight years later on a Vida 3T. SwiftSight-Brain found 32 lesions in the follow-up that were absent from the baseline, most of them in the deep white matter. Seven lesions that were already present, all periventricular, had grown, and their combined volume had more than tripled (+225.9%).

Age 31

Siemens Skyra 3T

Age 39

Siemens Vida 3T

Lesion dynamics

Across field strengths and scanner models

Retrospective case · Female patient, age 38–39 Siemens MAGNETOM Sola 1.5T and Vida 3T

This patient's annual follow-ups were acquired on two Siemens scanners that differ in both model and field strength. Most lesions remained stable, and the largest lesion, in the parietal lobe, grew slightly. SwiftSight-Brainclassified that lesion as enlarging. The conventional solution classified it as stable.

Age 38

Siemens Sola 1.5T

Classified as enlarging

Classified as stable

Across vendors

Retrospective cases · Philips 3T baseline, Siemens 3T follow-up

Vendor differences in noise and resolution can degrade segmentation accuracy. In four more patients with a Philips baseline and a Siemens follow-up, 11 months to more than three years apart, SwiftSight-Brain lesion quantification held steady despite the change of vendor.

Female Age 42–43 Feb 2019 – Nov 2020 Coronal view

Age 42–43

Feb 2019 – Nov 2020

Coronal view

Female Age 24–27 Jan 2020 – May 2023 Axial view

Age 24–27

Jan 2020 – May 2023

Axial view

Male Age 26–27 Jul 2020 – Jun 2021 Sagittal view

Age 26–27

Jul 2020 – Jun 2021

Sagittal view

Male Age 47–48 Feb 2020 – Jan 2021 Axial view

Age 47–48

Feb 2020 – Jan 2021

Request a Clinical Review

Our clinical team can provide in-depth clinical and technical presentations on SwiftMR.

Select publications

SwiftMR Information Security Whitepaper

Park J, et al. AIRS Medical

This white paper outlines AIRS Medical's information security practices for SwiftMR, and covers data workflow, encryption, disaster recovery, AI governance, and certifications. Cloud-based SwiftMR removes sensitive PHI from the DICOM tags as part of the de-identification process before transmitting the data to the SwiftMR Cloud Server.

SwiftSight-Brain Consistency Whitepaper

Jung W, Suh J, Yang S, Chung I, Kim D, Yang J

This white paper quantifies the measurement consistency of SwiftSight-Brain, benchmarked against an established open-source reference, along four axes of variation a deployed volumetry tool encounters in practice: repeated acquisition on a fixed scanner, a change in magnetic field strength (1.5T versus 3T), a change of scanner hardware, and a change in in-plane spatial resolution.

Leveraging a vendor-neutral deep learning reconstruction algorithm to reduce scan time and enhance image quality in T2-weighted liver MRI

Jeong B, Kim SS. Scientific Reports. 2026. doi:10.1038/s41598-026-60802-5.

In 86 patients scanned at 3T, accelerated T2-weighted liver imaging reconstructed with SwiftMR in a single breath-hold had higher SNR and sharpness than the routine two-breath-hold protocol, with lower spatial mismatch scores.

SwiftSight-Brain Normative Database Whitepaper

This white paper details the reference population and statistical methodology behind SwiftSight-Brain volumetric analysis, including a novel shifted-softplus normative model for age- and sex-adjusted assessment. The database draws on 3D T1-weighted MRI across major scanner vendors and field strengths to deliver stable, clinically interpretable percentile and z-score comparisons.

Vendor-neutral deep learning reconstruction of dynamic contrast-enhanced prostate MRI: image quality improvement and preservation of diagnostic performance

Lee YJ, et al. European Journal of Radiology. 2026;203:113020.

Across 173 patients, SwiftMR-reconstructed DCE series improved image noise, capsule sharpness, and lesion conspicuity for both readers, while PI-RADS diagnostic performance was preserved.

Denoising preclinical MRI with vendor-neutral deep learning-based image reconstruction

Oki T, et al. Journal of Neuroscience Methods. 2026;433:110813.

Preclinical study in six mice scanned at 4.7T, testing a vendor-neutral reconstruction trained only on human clinical MRI. SNR and CNR rose with each denoising level and sharpness improved, with fine anatomical structures preserved.

Automated estimation of perivascular space and brain morphometry from deep learning-reconstructed three-dimensional T1-weighted MRI: comparison with the conventional technique

Yang SJ, et al. Journal of Neuroradiology. 2026;53:101550.

In 240 participants plus an external dataset of 63, SwiftMR-reconstructed 3D T1-weighted images yielded greater cortical thickness and volume measurements and roughly halved perivascular space volume estimates, attributed to removal of noise that conventional images mislabel as perivascular space.

Deep Learning Reconstruction on Quantitative Analysis in Brain Tumors With Diffusion-Weighted Imaging and Dynamic Susceptibility Contrast Imaging

Cheong EN, et al. Journal of Magnetic Resonance Imaging. 2026.

ADC, CBV, CBF, and MTT derived from DWI and DSC perfusion showed no significant difference between original and SwiftMR-processed images across three enhancement levels, with high repeatability, while SwiftMR substantially reduced noise in the DSC time series.

Prospective evaluation of artificial intelligence in lumbar spine MRI workflow: from deep learning-enhanced accelerated acquisition to simultaneous vision-language model-based automated report generation

Park J, et al. European Journal of Radiology. 2026;196:112695.

A two-institution prospective study of 70 patients found SwiftMR-enhanced accelerated lumbar spine MRI diagnostically interchangeable with conventional imaging at approximately 80 to 86% shorter acquisition time, with higher SNR and CNR, and automated reporting agreeing closely with radiologist consensus.

A multinational study of deep learning-based image enhancement for multiparametric glioma MRI

Park YW, et al. Scientific Reports. 2025;15:32857.

Multinational retrospective study of 294 glioma patients at NYU, Severance, and SNUH. SwiftMR-enhanced T2-weighted, T2 FLAIR, and postcontrast T1-weighted images had higher SNR and CNR, better reader-rated image quality and lesion conspicuity, and fewer truncation artifacts.

Deep learning-based image enhancement for improved black blood imaging in brain metastasis

Oh G, et al. European Radiology. 2026;36(2):928-938.

This study evaluated SwiftMR for improving the image quality and diagnostic performance of 3D contrast-enhanced T1-weighted black blood MRI for brain metastases. Image quality and noise improved, and sensitivity increased, particularly for lesions smaller than 5 mm, while specificity remained comparable.

Deep learning-based reconstruction for accelerated cervical spine MRI: utility in the evaluation of myelopathy and degenerative diseases

Koo SJ, et al. American Journal of Neuroradiology. 2025;46(4):750-757.

In 50 patients, accelerated cervical spine MRI reconstructed with SwiftMR reduced acquisition time by 47 to 48%, had higher or similar image quality compared with the conventional protocol, and raised average sensitivity for neural foraminal stenosis from 65% to 73%.

Accelerated intracranial time-of-flight MR angiography with image-based deep learning image enhancement reduces scan times and improves image quality at 3-T and 1.5-T

Jeon YH, et al. Neuroradiology. 2025;67(5):1203-1213.

In 129 patients, 40% accelerated TOF-MRA with SwiftMR showed significantly higher SNR, CNR, contrast ratio, and vessel sharpness than conventional TOF-MRA, with improved visibility of small- and medium-sized vessels at both 3T and 1.5T.

Prospective Evaluation of Accelerated Brain MRI Using Deep Learning-Based Reconstruction: Simultaneous Application to 2D Spin-Echo and 3D Gradient-Echo Sequences

Choi KS, et al. Korean Journal of Radiology. 2025;26(1):54-64.

Across 150 participants scanned on 3T systems from three vendors, SwiftMR reduced brain MRI scan time by a mean of 39.4% while improving overall image quality, structure delineation, and artifacts. Volumetric quantification showed no significant differences between conventional and SwiftMR-processed accelerated 3D images.

Deep learning-based super-resolution and denoising algorithm improves reliability of dynamic contrast-enhanced MRI in diffuse glioma

Lee J, et al. Scientific Reports. 2024;14:25349.

In 306 patients with diffuse glioma, SwiftMR-enhanced DCE-MRI had higher SNR and CNR, more reliable arterial input function measurements, and improved differentiation of WHO grades based on the extravascular-extracellular volume fraction (AUROC 0.88 vs. 0.83).

Reliability of brain volume measures of accelerated 3D T1-weighted images with deep learning-based reconstruction

Jung W, et al. Neuroradiology. 2025;67:171-182.

Brain volumes and age- and sex-matched normative percentiles from SwiftMR-reconstructed accelerated 3D T1-weighted scans agreed closely with conventional scans (volume ICC > 0.96, percentile ICC > 0.89), supporting acceleration of up to 75%.

Seo M, et al. Scientific Reports. 2024;14:18983.

Across 86 lesions, SwiftMR improved vessel wall image quality and SNR, increased reader confidence in lesion detection, and raised sensitivity for identifying culprit plaques by intraplaque hemorrhage from 23.5% to 70.6% at equivalent specificity (94.3%).

Clinical Implications and Use Cases of SwiftMR

Yang K, et al. AIRS Medical

This white paper compares accelerated MR images processed with SwiftMR to the standard of care in a multi-center, multi-reader study of 184 exams across neuro, musculoskeletal, and body imaging. Eighteen board-certified radiologists from six subspecialties scored the SwiftMR-processed images higher than the standard of care in 74.7% of cases.

Demonstrating the Economic Value of SwiftMR

AIRS Medical

This white paper models the economic value of SwiftMR for a hypothetical U.S. outpatient imaging center with one MRI scanner. In the base case, increased scan volume produced $271,790 in additional revenue over one year, an ROI of 453%, and a payback period of 3 months.

SwiftMR Improves Image Quality and Diagnostic Performance of 4D Time-Resolved, Contrast-Enhanced MRA for Ischemic Stroke Imaging

Kim BK, et al. Diagnostics. 2024;14(11):1199.

This study applied SwiftMR to 4D time-resolved contrast-enhanced MRA in 520 patients evaluated for ischemic stroke or intracranial vessel stenosis. Processed images showed better image quality, sharpness, vascular conspicuity, and SNR, with a better confidence level and shorter decision time for large vessel occlusion.

Introduction to the Core Algorithm of SwiftMR

Jeong G, et al. arXiv. 2024. arXiv:2405.03684.

This paper introduces an all-in-one deep learning framework for MR image reconstruction, in which a single DICOM-based model enhances image quality across multiple aspects of k-space sampling. The algorithm serves as the core of SwiftMR and is compatible with various vendors, pulse sequences, scan parameters, and anatomical regions.

Optimizing SwiftMR Protocols

Jeong G, et al. AIRS Medical

White paper by Geunu Jeong, MD, Head of SwiftMR Research at AIRS Medical, on protocol-setting strategies for scan time reduction, spatial resolution improvement, and SNR improvement. Its guiding principle: adjust scan parameters by willingly sacrificing either SNR or spatial resolution, and SwiftMR will compensate for the loss.

SwiftMR Enhances Brain Epilepsy Image Quality

Suh PS, et al. Korean Journal of Radiology. 2024;25(4):374-383.

Retrospective study of 117 patients evaluating 1.5-mm coronal T2-weighted MRI with deep learning-based reconstruction (SwiftMR) for temporal lobe epilepsy. Pooled diagnostic accuracy on the same 1.5-mm images was 80.8% with SwiftMR versus 73.1% without it.

Questions about a study? Contact bd@airsmed.com.

Ready to see SwiftMR® at your institution?

Our clinical team can provide in-depth presentations on SwiftMR — including protocol-specific performance data for your scanner and field strength.